Cursor gets a massive update; OpenAI cozies up to both AWS and Azure
Get bigger weekly updates! For a limited time, you can grab a month-long trial of our paid newsletter which includes an extended weekly updates (six additional stories this week!) and some other benefits.
last week’s top stories
🤖 Cursor gets a massive update. AI coding assistant Cursor launched version 2.0, introducing a new model called Composer that runs code generation 4x faster than peers (most tasks finish in under 30 seconds). The update revamps the IDE with a multi-agent interface (among many other changes); developers can deploy up to 8 AI agents in parallel to tackle different coding tasks simultaneously, aided by built-in tools like an in-editor browser for testing code autonomously. Read more
☁️ AWS and OpenAI ink $38B deal for AI cloud power. Amazon Web Services and OpenAI announced a multi-year strategic partnership that gives OpenAI access to hundreds of thousands of Nvidia GPUs on AWS to run and scale its AI models. Under the $38 billion, 7-year agreement, OpenAI can rapidly expand compute across AWS (scaling to “tens of millions of CPUs”) to support ChatGPT and future models, supplementing its existing Azure cloud setup. Read more
🤝 OpenAI and Microsoft renew partnership with new Azure terms. OpenAI’s biggest backer Microsoft has restructured its deal with the AI lab as OpenAI transitions to a capped-profit public benefit corporation. In the new agreement, Microsoft will hold about 27% of OpenAI (worth $135 billion) and retains exclusive rights to OpenAI’s models on Azure through 2032, but drops its previous veto on OpenAI using other clouds. Notably, OpenAI agreed to spend an extra $250 billion on Azure over time, yet can now jointly develop some non-Azure products and even provide open-source “open weight” models. Read more
💰 OpenAI floats ambitious $1 trillion IPO plan. OpenAI has begun early preparations for an IPO that could value the company at up to $1 trillion, according to sources. The ChatGPT-maker is reportedly eyeing a filing as soon as late 2026, potentially raising ~$60 billion or more, though plans could shift with market conditions. OpenAI’s recent corporate restructuring (spinning off a capped-profit arm) and surging revenue growth have set the stage for a blockbuster offering, one that would make it among the largest IPOs in history if the $1T valuation materializes. Read more
🔥 New rumored details emerge from OpenAI’s boardroom coup saga. A new exposé reveals that after CEO Sam Altman’s surprise firing in 2023, OpenAI’s board frantically explored a merger with rival Anthropic to steady the company. Testimony from co-founder Ilya Sutskever indicates merger talks began within days of Altman’s ouster; a sign of the board’s desperation amid employee uproar and investor pressure (Microsoft intervened to get Altman reinstated). Read more
✂️ Amazon to slash 30k jobs as AI automates work. Amazon is planning its biggest corporate layoff ever (up to 30,000 jobs; ~10% of office staff) as it cuts costs and incorporates more AI-driven efficiency. CEO Andy Jassy has cited the rise of AI tools automating repetitive tasks as a factor enabling these cuts, on top of a drive to reduce bureaucracy. Read more
🎵 Universal, Udio settle on AI music. Universal Music Group settled a landmark lawsuit with AI song-generator Udio and agreed to jointly launch a licensed AI music platform. In the deal, Udio will license Universal’s catalog for AI remixing and mashups, but Udio had to disable all song downloads, sparking outrage among users who suddenly can’t save the AI-generated tracks they created. The settlement (the first of its kind) is a blueprint for how AI music tools might operate legally. Read more
⚡ Extropic claims 10,000x AI energy efficiency leap. Startup Extropic showed off a prototype “thermodynamic” AI chip that it says can run generative models with orders of magnitude less energy than today’s GPUs. The new processors, called Thermodynamic Sampling Units (TSUs), don’t use binary 1/0 bits; instead they harness random electron fluctuations (“p-bits”) to directly model probabilities; a physics-inspired approach to computing AI tasks like diffusion models. Read more
🖌️ Adobe boosts Creative Cloud with generative AI everywhere. At Adobe MAX 2025, Adobe rolled out a slew of AI features for creators across Photoshop, Illustrator, Premiere and more. A new Firefly Image Model 5 (4x higher resolution) now powers generative fills and upscaling in Photoshop, and Adobe integrated partner models (Google, OpenAI, etc.) directly into apps to give users choice of AI engines. The company also introduced “AI Assistants” in tools like Express and Photoshop and a video-to-soundtrack generative music model. Read more
🆘 ChatGPT data reveals widespread user mental distress. OpenAI disclosed new findings that over 1 million ChatGPT users per week send messages indicating suicidal intent or planning. The internal analysis (one of the first public stats on this issue) also estimated 0.07% of weekly users (~560k people) show possible signs of psychosis or mania during chats. OpenAI shared the sobering data as it rolled out improved crisis response features in its GPT-5 model (like offering hotline info and limiting harmful content). Read more
🔞 Character.AI will block minors from chatting with bots. Citing growing safety concerns, popular chatbot platform Character.AI announced it will bar users under 18 from open-ended AI chats starting next month. Teen users will still be allowed to use the service’s AI characters in structured ways (like generating videos or stories), but unrestricted back-and-forth chatting with the AI is being disabled for minors. Character.AI says it’s deploying in-house age-detection algorithms and verification prompts to enforce the policy. Read more
🧪 AI Research of the Week
Emergent introspective awareness in large language models
From Anthropic
Jake’s Take: Anthropic tested whether models can notice what is happening inside them by inserting concept signals into hidden layers and asking the model to report what changed. Claude Opus 4/4.1 correctly identified injected concepts about 20% of the time when the signal landed in the right layer and strength; it also separated these internal “thoughts” from the input text and judged whether a previous output matched its own prior intention. Models could even adjust internal representations on command (e.g., “think about X”).
This type of research helps push model interpretability from just reading outputs to causally probing their internals, giving Anthropic (and others) a concrete way to test, steer, and verify model intent (good news for safer deployments, reproducible evals, and policy-grade audits).
and then, even more news…
📖 Elon Musk launches Grokipedia as an AI-run Wikipedia. X.AI, Elon Musk’s AI startup, has released Grokipedia, an AI-generated online encyclopedia meant to rival Wikipedia. The site launched with ~885,000 articles created by Musk’s new Grok model (often adapted from Wikipedia entries) and pitches itself as seeking “the truth, the whole truth and nothing but” (a thinly veiled jab at Wikipedia’s alleged biases). Grokipedia’s debut has stirred controversy: early reviewers found politically skewed content on some pages, and Wikimedia’s response noted that Grokipedia has minimal citation and oversight compared to Wikipedia’s community-edited, source-backed articles. Read more